Title: FigSIM: A New Dataset for Analyzing Figurative Language and Fine-Grained Suicide Severity in Suicide Memes

Abstract

Despite their growing prevalence on social media platforms, suicide memes—visual content used to articulate suicidal ideation or discuss suicide-related topics—remain largely unexamined and carry significant risks of harm. There is a critical demand to comprehend the nuances of these memes and to implement effective content moderation protocols that shield users from dangerous material. However, the development and assessment of automated moderation systems are currently hindered by a lack of annotated datasets dedicated to suicide memes.

To address this gap, we present FigSIM, the inaugural dataset specifically constructed for the fine-grained analysis of such content. Comprising 1,049 memes, FigSIM provides detailed annotations across three dimensions: (1) nuanced levels of suicide severity, (2) the presence of figurative devices such as metaphors, and (3) specific suicide-related elements, including depictions of suicide methods.

We evaluated 16 unimodal and multimodal models on three distinct tasks: detecting figurative language, assessing suicide severity, and identifying suicide-related content. Our findings indicate that suicide memes present distinct difficulties for both computational modeling and human moderation. Notably, our analysis uncovered specific biases, including a tendency to underestimate higher levels of suicide severity, particularly within memes that rely on figurative language. The complete dataset, along with the data splits utilized in our analyses, is publicly accessible.

Content Warning: This paper includes material related to suicide that may be distressing or triggering for some readers.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC